# Memory Search

> Search and retrieve information from the Appie brain — daily logs, project notes, decisions, topics, and long-term memory files stored in ~/.hermes/memory/. Use this skill whenever you need to recall past conversations, decisions, project context, personal preferences, or anything discussed previously.

- Skill: `s3yed/memory-search` (Agent Skill)
- Install (CLI): `npx skillmds@latest add s3yed/memory-search`
- Raw SKILL.md: https://api.skillmd.com/api/skills/s3yed/memory-search/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- License: MIT
- Author: S3YED (https://skillmd.com/u/s3yed)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/s3yed/memory-search

---


# Memory Search

Search the Appie brain — daily logs, project notes, decisions, topics, and long-term memory.

## When to Use

- User asks about past decisions, conversations, or context
- User says "remember when...", "what did we decide about...", "check my notes on..."
- You need project context, client info, or infrastructure details
- Looking up dates, events, or historical information

## Memory Structure

All files are in `~/.hermes/memory/`:

- **Daily logs**: `YYYY-MM-DD.md` — raw daily notes
- **Topics**: `topics/*.md` — subject-specific knowledge
- **Projects**: `projects/*.md` — project documentation
- **Decisions**: `decisions/*.md` — recorded decisions
- **Infrastructure**: `infrastructure/*.md` — server/fleet notes
- **Plans**: `plans/*.md` — future plans
- **Research**: `research/*.md` — research notes

Top-level brain files in `~/.hermes/`:
- `MEMORY.md` — curated long-term memory (read this first)
- `USER.md` — info about the human (Seyed)
- `TOOLS.md` — tool configuration and fleet SSH details
- `IDENTITY.md` — agent identity

## Search Commands

### Full-text search across all memory files
```bash
grep -ril "SEARCH_TERM" ~/.hermes/memory/ ~/.hermes/MEMORY.md ~/.hermes/USER.md ~/.hermes/TOOLS.md 2>/dev/null
```

### Search with context (shows surrounding lines)
```bash
grep -rin -C 3 "SEARCH_TERM" ~/.hermes/memory/ 2>/dev/null | head -60
```

### List recent daily logs
```bash
ls -t ~/.hermes/memory/????-??-??.md 2>/dev/null | head -10
```

### Read a specific daily log
```bash
cat ~/.hermes/memory/YYYY-MM-DD.md
```

### Search by date range (e.g., March 2026)
```bash
ls ~/.hermes/memory/2026-03-*.md 2>/dev/null
```

### Search topics
```bash
ls ~/.hermes/memory/topics/ 2>/dev/null
grep -ril "SEARCH_TERM" ~/.hermes/memory/topics/ 2>/dev/null
```

### Search projects
```bash
ls ~/.hermes/memory/projects/ 2>/dev/null
grep -ril "SEARCH_TERM" ~/.hermes/memory/projects/ 2>/dev/null
```

### Advanced: fuzzy search with script
```bash
python3 ~/.hermes/skills/productivity/memory-search/scripts/search.py "QUERY" [--limit 10] [--recent 7]
```

## Quick Reference

Always start with `MEMORY.md` for curated long-term context, then search daily logs for specifics.

## Workflow

1. Read `~/.hermes/MEMORY.md` for high-level context
2. If more detail needed, search with `grep -ril` across memory/
3. Read the specific files that match
4. Synthesize and answer

## Project PRD / "last wishes" recall pattern

When the user asks whether you remember a PRD, brief, handoff, or their "last wishes":

1. Ground the answer before claiming memory. Search both long-term memory and the likely project directory for `PRD`, `Handoff`, `README`, `Open decisions`, `TODO`, `wishes`, and the project/client name.
2. Prefer canonical project artifacts over chat fragments: `docs/PRD*.md`, `HANDOFF.md`, `README.md`, `AGENTS.md`, `PROJECT_STATE.md`, and deployment manifests.
3. If multiple projects match, use the strongest local context and say which project you verified. Ask only if the ambiguity changes the answer.
4. Separate confirmed facts from inferred "last wishes". Example shape: "Verified PRD says X. Handoff/open decisions say Y. I do not see a separate file literally named last wishes."
5. Keep the reply short unless the user asks for the full doc. For Seyed, answer with the remembered decisions and blockers, not a tool-by-tool narrative.